How much GPS data do we need? - Home | Connected...

1

How much GPS data do we need?

Anthony Patire Matthew Wright

Boris Prodhomme Alexandre Bayen

1

Motivation

• Emergence of mobile internet and proliferation of smartphones

• Location data increasingly collected by private companies

• Huge opportunities for transportation organizations – Performance measures – Incident response – Situational awareness – Coordination and management

1

Hybrid Data Project Background

• Focus on probe data – commercially available, unaggregated, GPS-

based, point-speed data – Fields include: ID, lat, long, timestamp, speed

• Understand practical challenges – Procurement – Real-time feeds – Filtering – Map-matching – Fusing with other data types

• Assess benefits

1

Outline

• Introduction

• One answer: how much data we need

• Overview of data and method

• Results

• Probe data quality

• Conclusion

1

Travel Time MAPE for combinations of loop and probe data

I-880 NB congested periods between 3/2 and 3/17, 2012

1

Outline

• Introduction



• Results


• Conclusion

1

Overview of Data

I-15, Victorville

I-880, Fremont

I-15, Ontario

• Bluetooth – Three two-week deployments of ten

detection units – Independent measure of travel times

• Probe data – Two vendors – March and April of 2012

• Loop data from PeMS

1

Overview of Method

Flows (Loops) Filter

Data Fusion

Velocity Map

Calculate travel times

and aggregate

Aggregated travel times

Velocities (Probes)

Bluetooth Travel Time

Measurements Filter Aggregate Aggregated

travel times

Compare

PIF

1

Aggregated Travel Times from Bluetooth

• For each 15-minute period, travel times between successive Bluetooth sensors are measured:

• 𝑨𝑨𝑨𝑨𝑨𝑨 𝑻𝑻𝑻𝑻 = 𝑻𝑻𝑻𝑻+𝑻𝑻𝑻𝑻𝑻𝑻

• Average travel times taken as ground truth and compared with outputs of data fusion

• For visualization purposes we calculate average velocities:

• 𝑨𝑨𝑨𝑨𝑨𝑨 𝑽𝑽𝑨𝑨𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝑽 = 𝒔𝒔𝑨𝑨𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝒔𝒔 𝑽𝑽𝑨𝑨𝒔𝒔𝒍𝒍𝑽𝑽𝒍𝒍𝑨𝑨𝑨𝑨𝑨𝑨 𝑻𝑻𝑻𝑻

15

min 15

min 15

min

BT 2

BT 3

BT 4

T1 T2

Time

Distance

BT Measurements Filter Aggregate Aggregated

travel times

1

Bluetooth Average Velocities

880 Northbound – Wed. March 7, 2012

1

Map-matched Probe Data Filtered Loop Data

Filtered and Map-Matched Data


Data Fusion

Velocity Map

Calculate travel times and aggregate


Velocities (Probes) PIF


1

Velocity Maps

Assimilated Loop Data Assimilated Probe Data


Data Fusion

Velocity Map





1

Average Velocities Flows

(Loops) Filter Data

Fusion Velocity

Map




Average Velocities from Loop Data

Average Velocities from Probe Data


1

Outline

• Introduction



• Results


• Conclusion

1

Definition: Mean Absolute Percentage Error (MAPE)

• 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 = 1𝑁𝑁𝑁𝑁

∑ ∑ 𝑇𝑇𝒎𝒎𝑽𝑽𝒎𝒎𝑨𝑨𝑽𝑽 𝑖𝑖,𝑗𝑗 −𝑇𝑇𝑏𝑏𝑏𝑏(𝑖𝑖,𝑗𝑗)𝑇𝑇𝑏𝑏𝑏𝑏(𝑖𝑖,𝑗𝑗)

𝑁𝑁𝑗𝑗=1

𝑁𝑁𝑖𝑖=1

– 𝑻𝑻𝒎𝒎𝑽𝑽𝒎𝒎𝑨𝑨𝑽𝑽 𝑽𝑽, 𝒋𝒋 = 𝑴𝑴𝑽𝑽𝒎𝒎𝑨𝑨𝑽𝑽𝑨𝑨𝒎𝒎 𝒂𝒂𝑨𝑨𝑨𝑨𝒂𝒂𝒂𝒂𝒍𝒍𝑨𝑨 𝑽𝑽𝒂𝒂𝒂𝒂𝑨𝑨𝑨𝑨𝑽𝑽 𝑽𝑽𝑽𝑽𝒎𝒎𝑨𝑨𝒔𝒔 𝒇𝒇𝑽𝑽𝒂𝒂 𝑽𝑽𝑽𝑽𝒎𝒎𝑨𝑨 𝑽𝑽, 𝒂𝒂𝑽𝑽𝒂𝒂𝒎𝒎 𝒔𝒔𝑨𝑨𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝒔𝒔 𝒋𝒋 – 𝑻𝑻𝒃𝒃𝑽𝑽 𝑽𝑽, 𝒋𝒋 = 𝑩𝑩𝑽𝑽𝑩𝑩𝑨𝑨𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝒍𝒍 𝒂𝒂𝑨𝑨𝑨𝑨𝒂𝒂𝒂𝒂𝒍𝒍𝑨𝑨 𝑽𝑽𝒂𝒂𝒂𝒂𝑨𝑨𝑨𝑨𝑽𝑽 𝑽𝑽𝑽𝑽𝒎𝒎𝑨𝑨𝒔𝒔 𝒇𝒇𝑽𝑽𝒂𝒂 𝑽𝑽𝑽𝑽𝒎𝒎𝑨𝑨 𝑽𝑽, 𝒂𝒂𝑽𝑽𝒂𝒂𝒎𝒎 𝒔𝒔𝑨𝑨𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝒔𝒔 𝒋𝒋 – 𝑵𝑵 = 𝒔𝒔𝑩𝑩𝒎𝒎𝒃𝒃𝑨𝑨𝒂𝒂 𝑽𝑽𝒇𝒇 𝑻𝑻𝟏𝟏 −𝒎𝒎𝑽𝑽𝒔𝒔𝑩𝑩𝑽𝑽𝑨𝑨 𝑽𝑽𝑽𝑽𝒎𝒎𝑨𝑨 𝒑𝒑𝑨𝑨𝒂𝒂𝑽𝑽𝑽𝑽𝒎𝒎𝒔𝒔 – 𝑴𝑴 = 𝒔𝒔𝑩𝑩𝒎𝒎𝒃𝒃𝑨𝑨𝒂𝒂 𝑽𝑽𝒇𝒇 𝑽𝑽𝒔𝒔𝑽𝑽𝑨𝑨𝒂𝒂 − 𝑩𝑩𝑻𝑻 𝒂𝒂𝑽𝑽𝒂𝒂𝒎𝒎 𝒔𝒔𝑨𝑨𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝑽𝒔𝒔𝒔𝒔

• A measure of the percentage error between travel time from Bluetooth sensors and the model’s result using available data, averaged over time and space

1

Velocity State Estimation: I-880

Position (12 mi)

Time (24 hr)

Loop data only

Loop data and

Probe data

Bluetooth only

Probe data only


1

Velocity State Estimation: I-880

Data MAPE or BTMAPE

Loop 21%

Probe 15%

Loop+Probe 10%

One standard deviation “noise floor” 19%

I-880 NB congested periods between 3/2 and 3/17, 2012

1

Site Description: I-15 Victorville SB

• Ten Bluetooth sensors placed at roughly 2 mile spacing

• No active loop detectors in place

• Extremely encouraging results

1

Results: I-15 Victorville

Bluetooth measured average velocities

Velocity Map from probe data


Ensemble Kalman Filter

Velocity Map



Velocities (Probes)

Bluetooth Travel Time

Measurements Filter Aggregate Aggregated

travel times

Compare

PIF

I-15 Southbound – Sunday April 22, 2012

1

Velocity Maps resulting from different subsets of available data

Position (18 mi)

Time (8 hr)

Use 9% of probe data

yields 36% MAPE


yields 15% MAPE


yields 13% MAPE


yields 24% MAPE

1

Outline

• Introduction



• Results


• Conclusion

1

Discussion of Data Quality

• Example vehicle trajectories illustrating sample rate and penetration rate

High Sample Rate Low Penetration Rate

Low Sample Rate High Penetration Rate

spac

e

spac

e

time time

1

Penetration Rate vs. Sample Rate (1)

• Probe Set A data – Higher sample rate – Lower penetration rate: 0.1% – 85th percentile flows: 7 veh/hour

• Probe Set B data – Lower sample rate – Higher penetration rate: 0.6% – 85th percentile flows: 35 veh/hour

Set A Sample Rates Set B Sample Rates

1

Penetration Rate vs. Sample Rate (2) MAPE vs Probe Points MAPE vs Probe Devices

On freeways, a high penetration rate of probes is better than having a high sample rate. For this particular application, probe A data is oversampled.

1

Outline

• Introduction



• Results


• Conclusion

1

How much GPS data do YOU need?

For travel time estimation on freeways, here is a menu of data solutions for a given MAPE requirement:

1

Future Work

• Consider arterials and freeways in a coordinated way

• Decision support tools for a TMC – Estimation of current traffic conditions – Prediction of future traffic conditions – Assess impacts of potential control

interventions

1

Please visit our exhibition booth

University of California, Berkeley PATH Connected Corridors Program Booth 3208

Anthony D. Patire [email protected]

How much GPS data do we need? - Home | Connected...

Documents

Transcript of How much GPS data do we need? - Home | Connected...